mantelmoritz/CorridorKey-Synthetic-Streamers

CorridorKey Synthetic Streamers

244 AI-generated livestream / facecam video clips, mostly in front of green, teal and blue screens, with per-frame alpha mattes and foreground colours computed by CorridorKey 2.0 at native resolution. Every clip was reviewed by eye.

Why this dataset exists

Real green-screen footage of streamers is hard to come by: public stock footage shows actors in studios, while a real facecam setup looks very different. Typical facecam conditions:

• cheap webcams or mirrorless cameras with noise and 4:2:0 chroma
• pull-up screens that are too small, wrinkled or unevenly lit
• RGB lighting that spills onto hair and headphones
• gaming headsets, boom-arm microphones, glasses, mugs, energy drinks and desk clutter
• people leaning in and out of frame, a second person stepping in

This dataset was generated to train and evaluate a real-time green-screen keyer for OBS: the CorridorKey Real-Time Edge Refiner, a tiny network that restores native-resolution hair detail on top of CorridorKey 2.0 when CorridorKey runs at reduced resolution for live use.

It is useful for:

• Training and evaluating video matting / keying models on facecam-style content (hair, headsets, microphones, glasses, transparent cups, motion).
• Distillation: the labels are the output of a large, slow offline keyer (CorridorKey 2.0 at full resolution), so they are a natural target for small real-time students.
• Robustness tests: dark, noisy, spill-heavy, partially covered and badly lit screens on purpose.
• Re-compositing: straight foreground plus alpha lets you place the subjects on any background or on new, procedurally generated screens (which gives exact ground truth for the re-composited frames).

What is in it

Clips	244 (207 train / 37 test), about 5 s each, 24 fps (one 30 fps clip), 1253 s in total
Resolution	1920×1080 (122 clips), 2560×1440 (121), 1280×720 (1)
Labelled frames	40 consecutive frames per clip (frames 12–51), 9,760 frames in total
Labels	alpha (16-bit PNG) and straight foreground colour (sRGB, lossless WebP) at the clip's native resolution, plus a person mask of the first labelled frame
Backgrounds	229 clips with a screen (164 green, 33 teal, 29 blue, 3 not classified), 15 clips in a real room without a screen

Generator models

All people, rooms and objects are synthetic. No real person was filmed. Any resemblance to real people is coincidental. The clips were generated from text prompts (included per clip) with these video models:

Model	Clips
MiniMax H3 (MiniMax)	121
Kling 4.0 (Kuaishou)	116
Seedance 2.5 (ByteDance)	4
Kling 3.0 Turbo (Kuaishou)	1
Google Veo 3.1 Lite (Google)	1
Wan 3.0 (Alibaba)	1

Series

Series	Clips	Content
pilot	7	first comparison of generator models on the same prompt
s1	28	curly / long / frizzy hair, headphones, microphones, many screen colours and lighting setups
s2	108	wider variety of people, hair styles, accessories, desks, lighting; 6 real-room clips
s3	101	harder cases: foreground objects (cups, cans, hands, props), more screen types, partially covered screens, people standing up and leaving the frame; 8 real-room clips

Layout

metadata.jsonl                     one row per clip (see below)
clips/<id>.mp4                     the generated video, unchanged
labels/<id>/alpha/<frame>.png      16-bit grayscale alpha, <frame> = absolute frame index in the clip (000012 ...)
labels/<id>/fg/<frame>.webp        straight (un-premultiplied) foreground, sRGB 8-bit, lossless; 0 where alpha = 0
labels/<id>/sam_hint/000012.png    SAM 3.1 mask of the first labelled frame (person, headphones, microphone)
images/                            figures of this card

metadata.jsonl fields:

Field	Meaning
file_name	path of the clip
id	clip id (<series>_<number>[_room]_<generator>)
series	pilot, s1, s2, s3
generator	video model that generated the clip
split	train or test (the test clips were never used for training the edge refiner)
background	screen or real room (no screen)
screen_colour	green, teal, blue; empty for real rooms and for 3 screen clips whose colour could not be measured reliably
width, height, fps, num_frames	video properties
label_first_frame, label_num_frames	labelled frame range (12, 40)
label_source	how the label was made (see below)
review_tags	what the review saw: props_kept, extra_person, rim_light, blurry_source
prompt	the full generation prompt

Loading a labelled frame

import json, cv2, numpy as np

row = json.loads(open("metadata.jsonl").readline())
cid, f = row["id"], row["label_first_frame"] + 5

# The labels were computed on frames decoded with ffmpeg (-pix_fmt rgb24). OpenCV's decoder uses a slightly
# different YUV->RGB conversion (about 1-2 / 255 mean difference); for exact pixels use ffmpeg:
#   ffmpeg -i clips/<id>.mp4 -vf "select=gte(n\,12)" -fps_mode passthrough -frames:v 40 -pix_fmt rgb24 -start_number 12 %06d.png
cap = cv2.VideoCapture(row["file_name"])
cap.set(cv2.CAP_PROP_POS_FRAMES, f)
frame = cap.read()[1][:, :, ::-1] / 255.0                                         # RGB, sRGB-encoded

alpha = cv2.imread(f"labels/{cid}/alpha/{f:06d}.png", cv2.IMREAD_UNCHANGED) / 65535.0
fg = cv2.imread(f"labels/{cid}/fg/{f:06d}.webp")[:, :, ::-1] / 255.0               # straight sRGB foreground

background = np.zeros_like(fg); background[..., 2] = 1.0                           # any new background
composite = fg * alpha[..., None] + background * (1.0 - alpha[..., None])           # (compositing in sRGB for brevity)

How the labels were made

1. Keying. Each clip was decoded to lossless frames and keyed with the official CorridorKey 2.0 offline pipeline (infer.py, native resolution, overlapping 40-frame windows).

2. What counts as foreground. Everything physically in front of the screen stays in the matte: the person, but also mugs, microphones and boom arms, chairs and a second person. That is what a keyer for streaming should do: a streamer wants to hold things up to the camera. Things that are not in front of the screen (a monitor at the edge of the frame, a desk edge, a wall next to a too-small screen) cannot be keyed by colour. In a real setup they are cropped away, so in these labels they are often kept as foreground (see images/samples.png, first row).

3. Real-room clips. Without a screen, CorridorKey alone keeps most of the room. For 11 of these clips the label was recomputed with a SAM 3.1 (Meta) mask of the first frame as CorridorKey's alpha hint (label_source = "SAM 3.1 hint on the first frame"). The hint only steers the first frame, so the rest of the matte still comes from CorridorKey. 4 more real-room clips were already keyed cleanly without a hint. The sam_hint mask is included for every clip, but it was only used as a hint for those 11 clips.

4. Edge despill. CorridorKey's raw foreground keeps some screen colour in fine, semi-transparent hair. For screen clips this colour was removed in the edge band only: where alpha is below 0.98 (fully applied below 0.68) the screen's dominant channel is limited to the larger of the other two. Solid areas, for example green clothing or coloured hair, are not changed. Real-room clips are not despilled.

5. Review. All 275 generated clips were inspected on contact sheets (3 frames composited over grey, plus a 1:1 crop of the busiest edge region over black, over white and as alpha). Clips were left out when the label was clearly broken:

   • very dark sources where the hair alpha became a smeared or blocky blob
   • screen-coloured glow or haze in the hair that stayed in the foreground
   • half-transparent, blocky keys on furniture, shelves or door frames that would teach a model wrong edges
   • real-room clips where even the SAM hint did not separate the person from the room
   • generator artefacts (headphones morphing over the face, objects popping in, lighting jumping between colours)

   31 clips were removed this way. A measurement of semi-transparent areas outside the person mask flagged further candidates; all of them were checked by eye, and most were real motion blur or fur, which is correctly semi-transparent.

Important limitations

• These are pseudo-labels, not ground truth. They are as good as CorridorKey 2.0 is on the clip. That is very good on clean screens, but the review was a visual check, not a correction. Remaining typical errors:

  • small holes in dark headphone ear cups
  • soft, slightly blurry hair at native resolution
  • off-screen areas (monitors, desk edges) kept as foreground, see above

  Use them as a strong teacher signal, not as a benchmark of absolute matting accuracy.

• Real-room clips are few (15). They are useful as robustness material, but this is not a background-removal dataset.

• Generated video artefacts: hair and fine detail are produced by video generators. They can show smearing, temporal shimmer and occasional anatomical errors. They are realistic enough for keying, but not real camera footage. The test split of a model trained here should always be complemented by real footage.

• Only 40 frames per clip are labelled. The full clip is included for context and temporal models.

Changes

• v2 (October 2026): every clip reviewed by eye, 31 clips removed. Foreground despilled in the edge band. Real-room labels recomputed with a SAM 3.1 first-frame hint. New fields screen_colour, label_source and review_tags replace label_notes. Per-clip sam_hint masks added.
• v1: 275 clips, raw CorridorKey 2.0 output.

License and attribution

• Dataset (clips, labels, metadata): CC BY-NC-SA 4.0. Non-commercial use only, share alike, attribution required.
• Labels made with CorridorKey 2.0 by Corridor Digital LLC (model designed and trained by Zachary Harris in collaboration with Niko Pueringer and Corridor Digital): https://huggingface.co/nikopueringer/CorridorKey2.0, https://corridorkey.com. CorridorKey 2.0 itself is licensed under the Corridor Key Licence (a copy is included as LICENSE_CorridorKey.txt). This dataset does not contain CorridorKey code or weights.
• Hint masks made with SAM 3.1 by Meta (SAM License). Only the masks are included, no SAM code or weights.
• This is an independent community project. It is not affiliated with, sponsored by or endorsed by Corridor Digital LLC.

Citation

@misc{corridorkey_synthetic_streamers_2026,
  title  = {CorridorKey Synthetic Streamers: AI-generated facecam green-screen clips with CorridorKey 2.0 mattes},
  author = {Mantel, Moritz},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/mantelmoritz/CorridorKey-Synthetic-Streamers}}
}